For simulations to have any relevance you need the
model to predict behavior. Thus you need to know
the behavior - the real behavior - at a minimum.
Then almost certainly you will have to fit the model
to the reality beginning with the parameters which
are measurable or well known and then further
adjusting the "fudge factors" that remain, until you
have a model that was realistic for some anecdote,
or better yet a sample set with some variety and
that variation pushed at the measurable-params so
your "fudging" (the nonmeasurable) can be shown
to have some range of realism (because in modeling
it is not uncommon for 7-1/2 wrongs to make a
"right enough" (according to the modeling dude,
who does not have to deal with the consequences
of unrealism - that's the circuit designer's lot).
Seems to be a lot of interest in predicting reliability
based on models. But proving always has to happen
sooner or later. These proof tests are good places
to mine for data.
Probably a lot of what you will find for parameter
extraction will be oriented to automated, high-$
tools because modern models have too many
and too non-physical, non-electrically-measurable
parameters. But if you already have a time-zero
fitted model and a series of device curves across
aging that you can see movement on, that should
be approachable with "manual" methods once you
understand the model and which params do what.
When in doubt, start with the data.